Causal Couture Week 6: Launching the Causal Intelligence Layer

Week 6: Transitioning to Causal Intelligence

[0:00]Week five completed Causal Couture's phase four decision intelligence system. Observed signals, trends, alerts, structured recommendations, and user-driven scenario simulation. Week six begins the next layer, causal intelligence. The goal is not to declare that causality has been proven. It is to build the technical and methodological foundation for more rigorous causal analysis. Causal Couture now separates three different analytical claims. An observed signal describes what is present in historical retail data.

Constructing the Temporal Causal Dataset with Date Gaps

[0:34]A scenario simulation asks how the existing analytical framework responds to a user-defined hypothetical change. It is not a forecast. An estimated causal effect is an observational estimate of how changing one factor may relate to a later outcome under explicit assumptions. These outputs may inform one another, but they are not interchangeable. The first causal question implemented for the platform is, "What is the causal effect of social engagement on

The Causal Diagram: Confounders vs. Mediators

[1:03]Subsequent product sales?" The project keeps its existing analytical grain. One row for each date and SKU. The treatment is social engagement for a product on the current date. The outcome is units sold for that same product on the following day. That creates a clearly ordered relationship. Engagement at time T followed by sales at time T plus one day. To support this question, a dedicated temporal causal

Estimating the Effect & Running Bootstrap Uncertainty

[1:30]Data set was created from the unified retail data. It includes current and prior engagement, current and prior sales, current inventory, next period sales information, validity indicators, and inventory constraint indicators. A correction was important here. Simply shifting a sales column by one row can mistake the next available record for the next day, even when dates are missing. The updated pipeline checks the actual next observed date for each SKU.

Diagnostic Assurance: Placebo and Variation Checks

[2:01]Next day sales are populated only when that date is exactly one calendar day later. That keeps the temporal treatment and outcome relationship honest. The first causal diagram makes the assumptions visible. Prior sales may influence both current engagement and later sales. Prior engagement may also carry forward. Inventory availability can constrain how many units are sold. SKU specific characteristics and calendar effects may influence both engagement and purchasing

The Stockout Diagnostic: Identifying Suppressed Demand

[2:31]Behavior. Web variables such as product views or view to cart rate require caution. They may occur after social engagement and before purchase, making them possible mediators rather than confounders. The pathway could be social engagement, then web behavior, then purchase. Automatically controlling for those variables could remove part of the effect being studied. The first effect estimator uses linear regression. The preferred adjustment structure includes current engagement, prior

Frontend Causal Dashboard & Multi-Specification Robustness

[3:01]Sales, prior engagement, current inventory when available, SKU fixed effects, and day of week effects. Inventory constrained observations can be excluded from the primary estimate when appropriate. So, periods with no available stock do not automatically distort the relationship. The output reports an estimated engagement effect, the observations used, adjustment variables, fixed effects, and model information. But, the interpretation remains limited. This is an observational estimate from uploaded retail data, not experimental proof of causation. A single estimate can create a false sense of precision. So, the system adds bootstrap uncertainty analysis. It repeatedly resamples the available observations, re-estimates the model, and examines the resulting distribution of engagement effect estimates. The platform can then report a mean estimate, a median estimate, a bootstrap standard error, and a percentile interval. This is a first-stage uncertainty diagnostic. Later work can examine how to account more explicitly for dependence among observations from the same SKU. The causal layer also includes a shuffled treatment placebo test. Here,

The observed engagement treatment is randomly shuffled and the model is estimated again. If a similarly large relationship appears after the original treatment structure has been destroyed, the result requires more scrutiny. If the shuffled result is weaker, that provides additional diagnostic reassurance. But a placebo test does not prove causal identification. The system also checks whether engagement contains meaningful variation by examining its range, distribution, and number of unique values. This is a basic global variation check, not formal proof of conditional positivity or overlap across every confounder group. Retail data adds another complication. Observed sales can be limited by inventory. Low observed sales do not necessarily mean low customer demand if a product was unavailable. The stock out diagnostic identifies normal inventory periods, low stock periods, inventory pressure, and complete stock outs. It evaluates where observed sales may be an unreliable representation of underlying demand and assigns a suppressed demand risk classification. The system does not recover or impute true latent

Demand. It identifies where inventory constraints make interpretation more difficult. A new evidence layer brings the diagnostics together. It considers whether the effect could be estimated, whether uncertainty information is available, whether the bootstrap interval excludes zero, how the placebo compares, whether usable treatment variation exists, and whether inventory suppression risk is acceptable. The resulting categories can describe stronger, moderate, weak, or insufficient observational support. These categories are diagnostic summaries. They are not probabilities that the relationship is causal. The system also compares alternative model specifications, minimal adjustment, inventory adjusted, SKU adjusted, and preferred models. The question is whether the direction of the engagement effect remains consistent, or whether it is sensitive to specification. To make the result easier to understand, the platform also translates the estimate to a business scale change. Instead of presenting only the coefficient for a full one-point increase in engagement rate, it calculates the estimated sales effect associated with a 10 percentage point increase in engagement rate. This

Improves interpretation without changing the underlying observational status of the estimate. These components are consolidated into one causal intelligence workflow, unified retail data set, temporal causal data set, adjusted effect estimation, bootstrap uncertainty, placebo test, treatment variation and overlap, stockout and suppressed demand diagnostics, alternative specification robustness, observational evidence summary, and finally, the causal intelligence response. This gives the front end one coordinated analysis rather than requiring it to orchestrate every diagnostic independently. Phase five was integrated into the existing fast API back end through the causal intelligence route. get/phase5/causalintelligence The existing phase one through phase 4 capabilities were preserved while the new causal layer was added. The front end now includes a dedicated causal intelligence experience. It presents the estimated engagement effect, effect direction, the 10% point interpretation, an uncertainty interval, observational evidence level, placebo behavior, treatment variation status, stockout risk, robustness across specifications, diagnostic interpretation, and methodological warnings. These outputs are shown in a business-facing structure rather than as an unstructured

API response. The updated back end and front end were tested locally. The application reloaded successfully, Swagger documentation was accessible, the phase 5 end point was exposed, and the front end loaded with the new causal intelligence additions. The central separation remains, observed signal is not scenario simulation. Scenario simulation is not an estimated causal effect, and an estimated causal effect from observational data is not experimental proof. By the end of week 6, Causal Couture has a working first causal analysis architecture. It defines a treatment and temporally subsequent outcome, constructs the causal data set, estimates an adjusted observational effect, reports uncertainty, runs placebo and variation diagnostics, identifies inventory-related suppression risk, compares specifications, summarizes evidence, and presents the analysis through the platform. But this is the foundation of causal intelligence, not the completion of causal inference. Next, the work moves towards stronger sensitivity and temporal validation, improved uncertainty estimation, dependence within SKU-level observations, broader robustness testing, and eventually additional retail causal questions.